No new material exists yet. Vals AI reports that a team of its AI agents, working with author Geby Jaff, used physics simulations to propose two crystals that might one day suit computer memory, and nothing in the post has been synthesised or measured by the agents or anyone else.
The first crystal, YBaMnFeO₅, is a compound the agents designed. Jaff writes that, as far as the team could find, nobody has made it or proposed it for this purpose. The second, KV[Cr(CN)₆], is a compound chemists first made in 1999, and the agents flagged it as a match after searching for existing candidates. Both results are calculations on a perfect, idealised crystal.
The physics is worth a minute, because it explains why anyone wants this. Electrons carry a built-in property called spin, which you can picture as a tiny compass needle that points one of two ways. Memory stores a bit by which way those needles point. In a fridge magnet the needles in the atoms all line up, so their effects add and a field leaks out into the room.
An antiferromagnet alternates instead: one atom points up, its neighbour down, and the effects cancel. From outside, the material looks unmagnetic. A memory chip built from it would not disturb its neighbours, so the parts could sit closer together. Jaff also notes such materials switch about a thousand times faster, a figure the post gives without a citation.
The catch is reading the data back. In an ordinary antiferromagnet, electrons of both spins turn up at the same energies, so a circuit cannot tell them apart. The post’s answer is a third class, called Luttinger compensated. The up atoms and down atoms balance exactly, yet they sit in different surroundings, so electrons with one spin end up at different energies than the other. Total magnetism stays at zero while the spins can still be told apart.
For memory use, the gap matters. Room temperature jostles electrons by roughly 26 millielectronvolts. The agents calculated that YBaMnFeO₅ keeps its spins separated across windows of 1.0 and 1.4 electronvolts, about 38 and 54 times that jostle by our arithmetic. For KV[Cr(CN)₆] they calculated 2.6 and 1.6.
The designed compound has a problem the agents found themselves. It needs manganese and iron atoms in a strict checkerboard, and their own simulation shows that pattern dissolving near 950 K. Oxides like this form at roughly 900 to 1300 °C, where atoms barely move at the lower end, so a real sample would probably come out scrambled and lose the effect. Jaff concedes it may be hard to make.
The 1999 compound is the stronger lead, and it is the one with a measurement behind it. The chemists who made it recorded magnetic order up to 376 K, above room temperature. Chromium sits at the carbon end of the cyanide links and vanadium at the nitrogen end, so neither metal can swap seats, the very thing the designed compound could not promise. The sample is the weak point. The only one ever made is a damp powder, with water trapped inside it, and it carried a small leftover magnetism. The two simulation methods disagree on the damage: the more accurate one says the spin separation survives, while the faster one says the hole-side window would end up under half its size. The band gap and spin separation themselves remain unmeasured.
Two cautions apply. This is a company publishing its own agents’ output, with no peer review, and the post does not say which steps humans chose or checked. Jaff does publish inputs, raw outputs, analysis code, a list of caveats, and independent re-runs on GitHub, so others can audit the numbers.
The more interesting claim is that the agents did not invent the best material; they recognised one chemists had already made. The post says a 2008 study had plotted the relevant data without remarking on it. That is a literature-mining result, and a modest one.
The decisive step is cheap by materials standards: remake KV[Cr(CN)₆], dry it, and measure it. Until a lab does, memory engineers have a prediction and no part to evaluate.
Based on reporting by Vals AI (vals.ai), written by Geby Jaff and published 4 October 2026.
